Free OCR Is Metered in Four Different Units
Every comparison of free OCR tools lists them side by side as if "free" meant the same thing. It does not, and the unit decides which tool survives your actual job:
- Units per month — Google Cloud Vision, where each page of a PDF counts as a separate image.
- Pages per month — Azure Document Intelligence.
- Requests per month, with a file size cap — OCR.space, generous on count and tight on size.
- No meter at all — the local tools, limited only by your machine.
There is a fifth consideration that outranks all of them: whether the document should leave your computer. Receipts, invoices, payslips, contracts and ID documents are exactly what people photograph and OCR, and they are exactly what should not be uploaded to a third-party service. That alone settles the choice for a lot of real work.
Local Tools: No Limit, No Upload
PowerToys Text Extractor Fastest for screenshots
Free, from Microsoft, and already a keystroke away on Windows. Text Extractor is part of PowerToys: press Win+Shift+T, drag a box over any region of the screen — including text inside an image or a paused video — and the recognised text lands on your clipboard. It runs locally using Windows OCR language packs, so nothing is uploaded and there is no quota.
Two honest caveats from Microsoft's own documentation. It only recognises languages whose OCR language pack is installed, which you check and install through PowerShell — Get-WindowsCapability -Online | Where-Object { $_.Name -Like 'Language.OCR*' } lists what is available. And Microsoft's own page recommends using the Snipping Tool instead for capturing screenshots, since recent Windows versions include text extraction there too. Microsoft also notes the output "may not be perfect, so you have to do a quick proof read."
Tesseract Unlimited, scriptable, free
The open-source OCR engine behind a great many "free online OCR" websites — which is worth knowing, because you can run the same engine yourself without the upload. Tesseract is Apache 2.0 licensed (its Leptonica image library is BSD 2-clause), recognises more than 100 languages out of the box, and has used an LSTM neural network engine since version 4, with version 5 the current stable line. It is a command-line program and a library, so there is no quota and no account.
The cost is setup: you install it, you point it at files, and you get plain text out. For a folder of 500 scanned pages that is a one-line loop and no bill. For one screenshot it is overkill.
The Cloud Free Tiers
Google Cloud Vision 1,000 units/month
Google's pricing page puts the free allowance at the first 1,000 units per month for both Text Detection and Document Text Detection. After that it is $1.50 per 1,000 units up to 5 million, dropping to $0.60 beyond.
The detail that catches people out: each page of a multi-page file counts as a separate image. A single 40-page scanned PDF is 40 units, so the free tier is 25 documents of that size — not 1,000 of them. Accurate and well documented, but plan in pages rather than files.
Azure Document Intelligence 500 pages/month
Azure's free F0 tier allows 0–500 pages free per month and, per its pricing page, lets you test all Document Intelligence features, though a footnote notes the free instance does not support premium features. That is a smaller allowance than Google's in raw count, but it is counted in pages directly, and Document Intelligence does more than plain text: layout, receipts, invoices and ID documents have dedicated prebuilt models.
Its paid per-1,000-page prices were not displayed on the pricing page when checked, so budget via Azure's calculator rather than trusting a number from a blog.
OCR.space API 25,000 requests, 1 MB each
The most generous count and the tightest size limit. The free API plan allows 25,000 requests per month with a rate limit of 500 requests per day per IP, a maximum file size of 1 MB, and a maximum of 3 pages per PDF. You need to register for a free API key. PRO is $30/month and PRO PDF $60/month, both raising the file limit to 5 MB, with PRO PDF handling 999+ pages.
One megabyte is the number to internalise. A photo from a modern phone exceeds it comfortably, so you will be compressing before sending — our guide to compressing images for free covers getting under that ceiling without destroying the text.
Google Drive and Docs 2 MB per file
The no-setup option most people already have. Right-click a file in Google Drive, choose Open with → Google Docs, and the image or PDF is converted to editable text. Google's support page lists the constraints: the file should be 2 MB or smaller, formats are JPEG, PNG, GIF and multi-page PDF, and the language is detected automatically across a long list including English, Spanish, French, German and Chinese.
The same page is unusually frank about quality requirements, and they are the best checklist in this article: text at least 10 pixels high, the document upright (rotate sideways images first), common fonts such as Arial or Times New Roman, and sharp images with even lighting and good contrast. It also warns what will not survive: bold, italics, font size and line breaks are likely kept, but lists, tables, columns, footnotes and endnotes are not likely to be detected.
Where OCR Still Fails
Those Google requirements generalise to every tool here, and they explain most bad results better than any accuracy benchmark.
Tables and columns are the big one. OCR reads text, not structure. A two-column page tends to come out interleaved, and a table becomes a stream of cell values with no rows. If your source is tabular, expect to rebuild the structure by hand — and if the output is destined for a spreadsheet, CSV to JSON helps once you have the rows separated properly.
Low resolution beats every engine. Ten pixels of text height is the practical floor. Enlarging a blurry screenshot afterwards does not add information; re-capturing at a higher resolution does.
The output always needs a clean-up pass. OCR text arrives with hard line breaks at the end of every visual line, stray hyphens from justified text, and the occasional invisible character. Our guide on removing line breaks fixes the first problem, AI Text Cleaner and Remove Extra Spaces handle the rest, and Word Counter is a quick sanity check that you captured the whole page rather than half of it. For a longer cleaning sequence, see how to clean text data online.
Proofread anything that matters. Microsoft says it plainly about its own tool, and it applies to all of them: a digit misread in an invoice total or a reference number is the kind of error that survives into your work unnoticed.
Comparison Table
| Tool | Free Limit | Runs Locally? | Watch Out For |
|---|---|---|---|
| PowerToys Text Extractor | Unlimited | Yes — Windows only | Needs OCR language packs; MS suggests Snipping Tool instead |
| Tesseract | Unlimited (Apache 2.0) | Yes | Command line; you do the setup |
| Google Cloud Vision | 1,000 units/month | No | Each PDF page = 1 unit; $1.50/1,000 after |
| Azure Document Intelligence | 500 pages/month (F0) | No | Free tier excludes premium features |
| OCR.space API | 25,000 req/month · 500/day per IP | No | 1 MB per file, 3 PDF pages; API key required |
| Google Drive → Docs | 2 MB per file | No | Tables, columns and footnotes not likely detected |
Which One Should You Use?
- A screenshot, right now, on Windows: PowerToys Text Extractor. Win+Shift+T, drag, paste. Nothing beats it for speed, and nothing leaves the machine.
- A folder of scans and no budget: Tesseract. Apache 2.0, 100+ languages, no meter — worth the afternoon of setup if this recurs.
- A one-off photo and no tools installed: Google Drive, opening the file with Google Docs. Keep it under 2 MB and upright.
- Many small images through a script: OCR.space's 25,000 free requests, as long as each file fits in 1 MB.
- Structured documents — invoices, receipts, IDs: Azure Document Intelligence's prebuilt models, within 500 pages a month. But think hard about the document type first.
- Anything confidential: local only. An invoice, a payslip or an ID photo sent to a free web OCR service has been handed to a third party, and the privacy reasoning is the same as for running AI models locally.